OpenDDE

OpenDDE

Predicts the 3D structure of proteins and other biological molecules down to each atom, for early-stage drug research. Runs from a command line on a GPU, a CPU or in Docker. A preview release.

Made by aurekaresearch on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: OpenDDE is an open-source foundation model for biological molecules. It works at the level of single atoms and folds several molecules together, so researchers can predict structures, design molecules and refine candidates in drug research.

It is a preview release. Command options, input and output fields and the released checkpoints can change from one version to the next.

Who it suits: Researchers who want an open model for structure prediction and who are comfortable with Python and a command line. A CUDA GPU on Linux helps, though a CPU works for small tests.

Industry | Who uses it | How they use it | Example
Drug research | Computational chemists | Predict how molecules fold together before lab work | A candidate protein chain saved in an input file
Antibody research | Antibody and antigen groups | Run the checkpoint tuned for antibody-antigen pairs | A prediction for one antibody and its target
Academic labs | Structural biology researchers | Try the model on short inputs on an ordinary computer | A nine-letter protein sequence run on a CPU
Machine learning research | Model developers | Install an editable copy of the code and experiment | A source checkout set up with the dev group

Anyone who needs a stable tool for production pipelines can skip it for now, since the release is a preview.

Where to use it:

• Linux with a CUDA GPU: the GPU install is meant for this setup.
• Any computer with Python 3.11 or newer: the CPU install works, only slower.
• Docker: a prebuilt image is available.
• Windows, macOS or Linux with Python: the Hugging Face tool downloads the model files.

Use case | Who | What they get back
Structure prediction | Computational biologists | Predicted structures written to an output folder
Antibody-antigen runs | Antibody researchers | Predictions from the checkpoint tuned for that pairing
Quick install check | Anyone setting it up | A small test prediction that needs no search databases

How to set it up: You need a few things in place first.

• Python 3.11 or newer
• uv, to create the virtual environment and pick the PyTorch backend
• A Linux CUDA environment for the GPU install, or any machine for the CPU install
• A folder for model data, set with the OPENDDE_ROOT_DIR variable (it defaults to a cache folder in your home directory)

1. Create a Python 3.11 virtual environment with uv and activate it.
2. Install OpenDDE with the GPU or the CPU backend.
3. Run the doctor command to check the install.
4. Get the checkpoint. The prediction command downloads the default one when it is missing, or you can fetch it yourself.

The GPU install:

uv venv --python 3.11
source .venv/bin/activate
uv pip install --torch-backend cu126 'opendde[gpu]'
opendde doctor

The CPU-only install:

uv venv --python 3.11
source .venv/bin/activate
uv pip install --torch-backend cpu 'opendde[cpu]'
opendde doctor

To prepare the data folder from a cloned source checkout:

export OPENDDE_ROOT_DIR=/path/to/opendde_data
bash scripts/download_opendde_data.sh

To skip the large search databases when you turn off the MSA, template and RNA MSA features:

bash scripts/download_opendde_data.sh --skip-search-database

To pull the model files with the Hugging Face tool, or the Docker image:

pip install -U huggingface_hub
hf download aurekaresearch/OpenDDE
docker pull aurekaresearch/opendde:v1

How to use it: Describe your molecules in a JSON input file, then run the prediction command with the input file and an output folder. This small input holds one protein chain.

[
  {
    "name": "tiny",
    "modelSeeds": [101],
    "sequences": [
      {
        "proteinChain": {
          "sequence": "ACDEFGHIK",
          "count": 1
        }
      }
    ]
  }
]

With every outside search turned off, this run needs no search databases:

opendde pred \
  -i tiny.json \
  -o ./output \
  -n opendde_v1 \
  --use_msa false \
  --use_template false \
  --use_rna_msa false \
  --sample 1 \
  --step 200 \
  --cycle 10

For the antibody-antigen checkpoint, name its file explicitly:

opendde pred \
  -i input.json \
  -o ./output \
  --load_checkpoint_path "$OPENDDE_ROOT_DIR/checkpoint/opendde_abag.pt"

Tool | What it does
opendde pred | Runs a prediction from an input file and writes results to the output folder
opendde doctor | Checks that the install is working

Example: Save the sample input with its single protein chain, then run the small prediction above. No search databases are needed because those features are off. Results are written to the output folder you named. On a CPU this run may be slow.

For newcomers and experts: If you are new, think of OpenDDE as software that estimates the 3D shape of a molecule from a description of its parts. Start with the CPU install, run the doctor command, then try the small prediction to confirm everything works.

If you know the field, this is an all-atom model built around co-folding, with a general checkpoint and one tuned on antibody-antigen data. Out of the box it infers in fp32 and picks its triangle kernels automatically, and a --seeds flag overrides the values in each job's modelSeeds field. Turning off MSA, template and RNA MSA trades extra input features for a much lighter data setup.

Describe the molecules in a JSON input file
Run opendde pred with an output folder
The checkpoint loads from the data folder
The model predicts the structure
Results are written to the output folder

Good to know: The points below cover what to expect.

• Preview status: options, input fields and checkpoints may change, and predictions may not repeat across releases.
• License: Apache 2.0.
• Cost to run: no API key or paid service is named. A GPU is optional, but CPU runs can be slow.
• Data it reaches: it reads checkpoints and runtime files from the data folder, and the prediction command can download missing default files. Local template and RNA searches need the large search databases.
• Not for: production pipelines, which the preview release is not meant for yet.

On Hugging Face, 13 November 2026

• Task: other
• Library: opendde
• Downloads in the last 30 days: 10,009
• Likes: 12
• License: apache-2.0
• Tags: opendde, biology, protein, protein-structure-prediction, drug-discovery, co-folding, all-atom, diffusion, other

Links

• Model: https://huggingface.co/aurekaresearch/OpenDDE
• Files: https://huggingface.co/aurekaresearch/OpenDDE/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download aurekaresearch/OpenDDE

Then follow the usage in its model card. Nothing in this kit runs by itself.